Professional Documents
Culture Documents
87
87
87
88
88
Introduction
1.1 Background
Long Term Evolution (LTE) is a project within the Third Generation Partnership Project (3GPP)
in order to improve the UMTS (Universal Mobile Telecommunications System) mobile phone
standard such that future requirements can be met. 3GPP is a collaboration agreement established in December 1998 and it is a co-operation between ETSI (Europe), ARIB/TTC (Japan),
CCSA (China), ATIS (North America) and TTA (South Korea). The good aspect of 3GPP is the
centralization of the standards, since a single organization for these technologies ensures global
interoperability1 .
3GPP standards are structured as releases, which incorporate several individual standards. Table
1.1 shows the latest releases and emphasizes some of the specications.
Developments currently done by 3GPP (Release 7 and above) are under the title UMTS Long Term
Evolution.
Version
Date
Description
Release 99
End of 1999
Release 5
2002
Release 6
End of 2004
In progress.
Release 7
Release 8
In progress
1.1.1
Release 5 and 6 incorporate enhancements such as HSDPA and HSUPA offering up to 10 Mbps
in downlink and 5.7 Mbps in uplink [23]. Using these enhancements, the 3GPP radio access
technology will be competitive for the next years. However, to ensure competitiveness for the next
1
http://www.3gpp.org/About/about.htm
CHAPTER 1. INTRODUCTION
several years and beyond, a long-term evolution of the 3GPP radio-access technology is currently
considered [23].
In particular, to enhance the capability of the 3GPP system to cope with the rapid growth in IP data
trafc [16], the packet-switched technology utilized within 3G mobile networks requires further
enhancement. A continued evolution and optimization of the system concept is also necessary in
order to maintain a competitive advantage in terms of both performance and cost [23].
Hence important parts of the long-term evolution include reduced latency, higher user data rates,
improved system capacity and coverage, and reduced cost for the operator. In order to achieve
this, an evolution of the radio interface as well as the radio network architecture is considered.
The present 3G network utilizes 5 MHz bandwidth for transmission between terminal also denoted
as user equipment (UE) and base station (NodeB), but if higher data rates are desired, future
spectrum allocations for LTE will evolve towards supporting wider transmission bandwidth. At
the same time, support for transmission bandwidths of 5 MHz and less than 5 MHz allows more
exibility in whichever frequency bands the system may be deployed. The main objectives of LTE
are the following [23]:
- User plane latency below 5 ms with 5 MHz or higher spectrum allocation. With narrower
spectrum allocation, latency below 10 ms should be facilitated.
- Scalable bandwidth up to 20 MHz, with smaller bandwidths covering 1.25 MHz, 2.5 MHz,
5 MHz, 10 MHz and 15 MHz for narrow allocations.
- Downlink peak data rates up to 100 Mbps.
- Uplink peak data rates up to 50 Mbps.
- Support for packet switched (PS) domain only.
- Up to 4 Tx-antennas at the NodeB and 4 Rx-antennas at the UE.
- Optimized performance for mobile speed of less than 15km/h, and high performance for
speeds up to 120km/h, and the connection should be maintained with mobile speeds up to
350 km/h.
In order to achieve the future requirements orthogonal frequency division multiplexing (OFDM)
is used in the physical layer for downlink purpose, which is specied in the current Release 7 [4].
Unlike HSDPA or HSUPA, high speed orthogonal packet access (HSOPA) is an entirely new air
interface system, unrelated to and incompatible with WCDMA.
10
11
CHAPTER 1. INTRODUCTION
12
Downlink
One of the main changes in the LTE system compared to 3G-UMTS is the physical layer. In third
generation systems, Wideband Code Division Multiple Access (WCDMA) is the most widely
adopted technology. A highlight of the characteristics of the UMTS before Release 7 is listed
below [12]:
- User information bits are spread over a wide bandwidth by multiplying the user data with a
spreading code. The use of variable spreading factor allows a variation of the bit rate.
- The bandwidth is 5 MHz. The chip rate used is 3.84 Mcps. A network operator can deploy
multiple 5 MHz bands to increase capacity.
- The frame length is 10 ms. During this phase, the user data rate is kept constant. However,
the data rate among the users can change from frame to frame.
In the LTE system, this will be very different. The new system will present an OFDM based
structure. The main aspects important for channel estimation in the physical layer are presented in
the following section.
FDM since the subcarriers are overlapping, but orthogonal. The frequency spacing between the
subcarriers fspace =
fs
NIFFT
8 [5]. In contrast to an OFDM transmission scheme, OFDMA allows multiple users to share the
available bandwidth. Each user is assigned a specic time-frequency resource referred as resource
block (RB). The fundamental principle of the Evolved UMTS Terrestrial Radio Access (E-UTRA)
is that the data channels are shared channels, i.e. for each transmission time interval (TTI) of 1
ms, a new scheduling decision is made at NodeB regarding which users are assigned to which
time/frequency resources during this transmission time interval.
In the LTE only packet-switched transmission is utilized [1]. OFDMA ts perfectly into packetswitched transmission, since different number of subcarriers (RBs) can be assigned to different
users, in order to support differentiated Quality of Service (QoS).
The scheduling is dynamic and performed for each subframe, hence the number of RBs can be
adjusted dynamically depending on the channel quality.
Figure 2.1: Frame structure in LTE [4]. A radio frame is divided into 20 slots of 0.5 ms each having 6 or 7
OFDM symbols. Two slots make one subframe, which corresponds to the minimum downlink TTI.
14
Transmission BW
1.25 MHz
2.5 MHz
5 MHz
10 MHz
15 MHz
20MHz
0.5 ms
15 kHz
Sampling frequency fs
1.92 MHz
3.84MHz
7.68 MHz
15.36 MHz
23.04 MHz
30.72 MHz
128
256
512
1024
1536
2048
75
150
300
600
900
1200
(4.69/144)6
7/6
Short
(s / sample)
Long
(4.69/9)6
(4.69/18)6
(4.69/36)6
(4.69/72)6
(4.69/108)6
(5.21/10)1
(5.21/20)1
(5.21/40)1
(5.21/80)1
(5.21/120)1
(5.21/160)1
(16.67/32)
(16.67/64)
(16.67/128)
(16.67/256)
(16.67/384)
(16.67/512)
Table 2.1: Downlink parameters for OFDM transmission. The occupied subcarriers {1, . . . , NBW } are
centered around the frequency f = 0 [4].
fs
NIFFT
= 15 kHz is used, and it is always constant, hence fs and NIFFT are proportional.
The downlink parameters for fspace = 7.5 kHz are not yet dened [5].
The number of OFDM symbols Nsym per slot depends on the length of the CP as described in
section 2.2. If 128-point IFFT and short CP is used, the rst 6 OFDM symbols have a CP of 9
samples and the last symbol a CP of 10 samples, such that the duration of the subframe of 0.5
ms is preserved. Not all subcarriers are occupied, in Release 7 [4] approximately 2/3 of the total
frequency band is used. According to technical specications in Release 8 [5] the number of used
subcarriers (here denoted as NBW ) can be varied. The values of NBW however are not specied. In
this project the values NBW are the same as in Release 7. Other downlink parameters than number
of FFT-points and sampling frequency are not yet determined, but the above assumption is used
for evaluation purpose in 3GPP, hence these parameters are also used in the project.
2.3.1
Mapping of Subcarriers
The subcarriers are mapped into the frequency spectrum as illustrated in Figure 2.2. According to
Table 2.1, NBW is 75/150/300/600/900/1200 when the transmission bandwidth is 1.25/2.5/5/10/15/20
MHz. Since the occupied subcarriers are centered around the frequency 0, half of the occupied
subcarriers are placed in the negative spectrum and the other half in the positive spectrum. Let us
denote the occupied subcarriers in the negative spectrum as {1, . . . , Nn } and in the positive spectrum as {Nn + 1, . . . , NBW }, where Nn (also shown on Figure 2.2) is 37/75/150/300/450/600
[4]. The unused carriers are placed at the edges of the spectrum such that the utilized bandwidth is
less than the specied bandwidth. This can be based on reducing the requirements for the analog
lters at the transmitter and receiver side.
15
Figure 2.2: Placement of occupied subcarriers [4]. NBW and Nn are the total number of occupied subcarriers and the number of carriers in the negative spectrum respectively.
resource block size is the same for all bandwidths, hence the number of available physical resource
blocks depends on the bandwidth. Depending on the required data rate, each UE can be assigned
one or more resource blocks in each transmission time interval of 1 ms. The scheduling decision is
done at the NodeB. The user data is carried on the Physical Downlink Shared Channel (PDSCH).
Downlink control signaling on the Physical Downlink Control Channel (PDCCH) is used to transport the scheduling decisions to individual UEs. The PDCCH is placed in the rst OFDM symbols
of a slot [5].
2.4.1
Modulation
According to the working assumptions for PDSCH in Release 8, the transmitted bits are modulated
using quadrature amplitude modulation (QAM). The available modulation schemes are 4-QAM,
16-QAM, and 64-QAM [4].
16
2.4.2
The downlink reference signal structure is important for cell search and channel estimation. Resource elements in the time-frequency domain are carrying the reference signal sequence, which
is predened for each cell.
Appendix ?? gives a detailed description of how the reference symbols are generated and on their
positions. In this section we focus on the main properties of the structure.
The reference symbols are placed in the rst OFDM symbol of one slot and on the third last OFDM
symbol [4]. The spacing between the reference symbols is always 6 subcarriers [4] and the norm
is always 1 no matter which modulation scheme is utilized for the data symbols as described in
Section 2.4.2.
In the LTE the NodeBs and UEs can have 2 or 4 antennas and when two or more transmitter antennas are applied, the reference symbols are transmitted such that they are orthogonal in space. The
orthogonality in space is obtained by letting all other antennas be silent in the resource element in
which one antenna transmits a reference symbol [5].
Figure 2.4 shows the positions of the reference symbols for transmission with two antennas as an
example. When antenna 1 transmits a reference symbol, antenna 2 is silent and vice versa. This
Figure 2.4: The reference symbol structure for one slot with 6 OFDM symbols using two antennas. Note
that only the used subcarriers are depicted. In this thesis we consider one antenna and makes use of the
reference symbol structure depicted for antenna 1.
thesis considers one antenna and makes use of the reference symbol structure depicted for antenna
1 on Figure 2.4.
The reference signal sequence also carries the cell identity. The reference signal sequence is generated as a symbol-by-symbol product of an orthogonal sequence (OS) ROS C3402 (3 different
sequences are predened) and a pseudo-random sequence (PRS) RPRS R3402 (170 different
sequences are predened). Each cell identity corresponds to a unique combination of one orthog17
2.4.3
Cell Search
During cell search, different types of information need to be identied by the UE such as radio
frame timing, frequency, cell identication, overall transmission bandwidth, antenna conguration, cyclic prex length. Besides the reference symbols, synchronization signals are therefore
needed during cell search. In E-UTRA (Evolved UMTS Terrestrial Radio Access) the synchronization acquisition and the cell group identier are obtained from different synchronization channels (SCH).
A primary synchronization channel (PSCH) for synchronization acquisition and a secondary synchronization channel (SSCH) for cell group identication have a pre-dened structure. They are
transmitted on the 72 subcarriers centered around subcarrier at frequency f = 0 within the same
predened slots (1st and 11th slot in one frame). PSCH and SSCH are however placed on the
second last and third last OFDM symbol respectively [5].
Hence cell search is always performed using the 72 central subcarriers independent of the overall
transmission bandwidth.
2.6.1
The binary source generator generates the signal randomly. The number of the generated binary
symbols depends on the modulation scheme, i.e. the number of bits per QAM-symbol (sec. 2.4.1)
and the number of subcarriers (sec. 2.3).
18
2.7. SUMMARY
2.6.2
Modulation
During modulation it is necessary to normalize the transmitted symbols in order to adjust the
signal-to-noise ratio. The normalization is achieved by scaling the symbols as listed in Table 2.2.
The bits are modulated using Gray labeling, which is depicted in Figure 2.7 for 16-QAM as an
Modulation
Knorm
4-QAM
1
2
1
10
1
64
16-QAM
64-QAM
Table 2.2: Normalization factor for M-QAM modulation schemes in E-UTRA downlink [5].
example.
2.6.3
Figure 2.8 shows an illustration of the mapping of symbols under IFFT-operation, which corresponds to the mapping described in section 2.3.1.
2.7 Summary
Important properties of the physical layer in the LTE downlink have been described. For the
purpose of channel estimation, one antenna at transmitter side and receiver side is considered. The
channel estimation can be achieved separately for each antenna since the reference symbols are
orthogonal in space. OFDMA is utilized as multiple access scheme in the downlink, where each
user is allocated one or several resource blocks and scheduling is performed for each subframe.
Based on specications in Release 7 [4] and on working assumptions in Release 8 [5] an OFDM
19
0000
0100
1100
1000
0001
0101
1101
1001
0011
0111
1111
1011
0110
1110
0.8
0.6
Quadrature
0.4
0.2
0
0.2
0.4
0.6
0.8
1
1
0010
0.5
0
InPhase
0.5
1010
1
Figure 2.7: 16-QAM constellation with the corresponding binary values after normalization.
20
Channel Model
3.1 Multipath transmission
Radio wave propagation can be described by multiple paths occurring due to reection (buildings,
trees, etc.) in the environment. When modeling a radio channel, only a nite number of paths is
considered to approximate the real environment, which is illustrated by Figure 3.1. The received
signal at the UE is a superposition of all paths. There are innite number of paths closely spaced
in time, but only a nite number is modeled. Equation (3.1) describes a model for the channel
impulse response,
h ( ) =
K1
X
k=0
k ( k ), where
(3.1)
K is the number of paths, k is the complex fading coefcient for a given path at delay k . In this
thesis a baseband representation of the channel is used, hence the effect of up-converting/transmit
lter and down-converting/receive lter is modeled by hT ( ) and hR ( ) respectively [28].
h( ) = hT ( ) h ( ) hR ( ),
(3.2)
Applying the sampling operation at rate 1/TC equation (3.2) is transformed into
h [k] = h [kTc ] , where
(3.3)
where the discrete time index k {0, ..., L 1} and L denotes the length of the sampled impulse
response. One channel model is the wide-sense stationary uncorrelated scattering (WSSUS). In
WSSUS model, the time-varying fading process is assumed to be wide-sense stationary random
process and the signal reections from the scatterings by different objects are assumed to be independent. The following parameters are often used to characterize a WSSUS channel:
- In order to describe the average delay of the channel the term delay spread, RMS is used.
- Coherence bandwidth f(c) gives an indication of how far apart, in frequency, the signal
has to be spaced to in order to achieve correlation of value c.
- Coherence time t(c) gives a measure of the time duration over which the correlation between two channel impulse response has a value c. The coherence time is related to the
Doppler spectrum which depends on the velocity of the UE.
- Doppler spread fDmax which indicates a maximum range of Doppler shifts.
The mentioned parameters are now described in detail and calculated for the signal structure of
the physical layer in the LTE downlink. Furthermore standard channel models will be described
in order to nd a suitable model for simulation purposes.
PK1
Pk k
Pk=0
K1
k=0 Pk
(3.5)
1
arccos(c),
2RM S
(3.6)
where f(c) is the minimal coherence bandwidth in which the autocorrelation of the channel
power delay prole, Rhh has the value c.
A multipath channel is characterized as frequency at or frequency selective in the following way:
Frequency at fading: Coherence bandwidth fc >> Bs Symbol bandwidth . The frequency
components of the signal would roughly undergo the same fading.
Frequency selective fading: fc Bs , with c = 0.5 is in both cases. The different frequency
components of the signal, which differ by more than fc will undergo different degrees of fading.
In order to avoid ISI the delay spread RM S should always be less than symbol duration Ts , such
that frequency at fading is achieved. The data rate however will be decreased if the symbol
duration is extended.
3.2.1
The position of the reference symbols in frequency domain is investigated. Section 2.4.2 describes
the positions in detail. The goal is to compare the spacing between the reference symbols with the
coherence bandwidth of the channel.
3.2.2
Coherence Bandwidth
The spacing in frequency between the subcarriers is known from Section 2.3,
fspace = 15 kHz.
(3.7)
According to Section 2.4.2 the reference symbols are placed at every 6th subcarrier:
fR = fspace 6 = 90kHz.
(3.8)
Using the uncertainty relationship described in Section 3.2 the bandwidth B(c) in which the channel is constant is dened for c = 0.9 and the value c = 0.5 for when the channel has changed.
Based on these assumptions the coherence bandwidths are now calculated:
B(0.5)
B(0.9)
1
1
arccos(c) =
arccos(0.5) = 256.4kHz
2RM S
20.65s
1
arccos(0.9) = 110.4kHz
20.65s
(3.9)
(3.10)
Equations (3.9) and (3.10) show that the spacing of the reference symbols in frequency approximately corresponds to the bandwidth in which the channel is constant, furthermore the channel is
estimated at least twice before the autocorrelation has the value 0.5.
23
3.3.1
Doppler shift
The main factor that affects the rate of fading is the mobility of the receiver relative to the transmitter. As a UE moves with velocity vUE relative to NodeB, it causes a Doppler shift fd which is
given by (3.11) for a singe path:
fd = fDmax cos(),
(3.11)
where is the angle of arrival of the received signal relative to the direction of the UE and
fDmax =
vUE
fc ,
c0
(3.12)
where fDmax is the maximum Doppler frequency , vUE is the velocity of the receiver, fc is the
carrier frequency and c0 speed of light.
The time varying impulse response of the channel can be expressed as an extension of Equation
(3.1).
h(t, ) =
L1
X
k=0
k ej2fd t ( k ),
(3.13)
where t is the time at which the channel impulse response is measured. The signal h(t, ) has a
band-limited spectrum, also denoted as Doppler spectrum. The spectrum is limited to the range
W = [Dmax , . . . , Dmax ] ,
where
Dmax = fDmax Ts ,
(3.14)
(3.15)
with Ts as the symbol duration and Dmax as the normalized maximum Doppler-shift. In mobile
radio channels, the maximum Doppler spread is used to characterize how fast in time the channel
changes. For this purpose the coherence time is calculated as [9]
t(c)
1
arccos(c),
2fDmax
(3.16)
where t(c) is the difference in time for which the autocorrelation in the time-domain has the
value c.
A channel is said to be time-invariant, in the sense that the channel appears as time invariant to the
transmitted signal.
Hence the terms are relative to the symbol duration: time selective fading occurs when the channel
changes within one symbol period Ts while time-invariant fading occurs when the channel is
constant within at least one symbol period.
24
symbol duration , the channel is considered to have time selective fading. In general, it is difcult
to estimate the channel parameters for a time selective channel.
In order to estimate a channel for a period of Ts it is of importance to be in the time-invariant case.
3.3.2
The duration of one slot Tslot is known from Section 2.3. There are Nsym = 6 OFDM symbols in
one slot, which results in a symbol duration
Ts =
Tslot
0.5 ms
=
= 83.3 s.
Nsym
6
(3.17)
Reference symbols are placed at each fourth OFDM symbol, but are not always found at same
subcarriers. Nevertheless the channel can be estimated after each third OFDM symbol and hence
the spacing of reference symbols in the time-domain is set to,
TR = 3Ts = 0.250 ms
(3.18)
Using the uncertainty relationship described in Section 3.2 the coherence bandwidth f(c) in
which the channel is constant is dened for c = 0.9 and the value c = 0.5 for when the channel
has changed.
A UE with a maximum speed of 120 km/h, has a maximum Doppler shift of 185.4 Hz. Based on
these assumptions the coherence times are now calculated
t(0.5)
t(0.9)
arccos(c) =
1
arccos(0.5) = 0.750 ms
2185.4Hz
2fd, max
1
arccos(0.9) = 0.323 ms
2222.15Hz
(3.19)
(3.20)
Equations (3.19) and (3.20) show that the spacing of the reference symbols in time is less than the
coherence time . Furthermore the channel is estimated at least three times before the autocorrelation of the channel estimates has the value 0.5.
(3.21)
Equation (3.21) is the autocorrelation of the time-varying channel frequency response, which is the
Fourier transform of (3.13) in the delay domain. The difference in time is denoted as t, while f
25
is the difference in frequency. Equation (3.21) is based on the fact that in mobile radio channels the
time-correlation function is independent from the frequency correlation function [13]. Moreover
different assumptions can be made for these correlation functions. The frequency correlation
function is the Fourier transform of the power delay prole of the channel, hence it is based on
the power delay prole. The time-correlation depends on the Doppler power spectral density. The
Doppler power spectral density can be assumed uniform or having Jakes spectrum [7]. In this
thesis the uniform spectral density is used. The constant spectrum can be expressed as,
(
1
2Dmax , W
Sh (, W) =
0
otherwise,
(3.22)
where Dmax is dened in equation (3.15) with as Dmax = fd Ts and W in (3.14). The discrete
correlation function with Rd (k) = R(kTs ), based on [27], is then expressed as,
1
Rd (k, W) =
ej2kDmax ej2kDmax
j2k|W|
sin(2kDmax )
=
k|W|
(3.23)
Pk = [k] =
Lk
PL1
Lk
k =0 e
LD =
, where
(3.24)
RMS
TC
(3.25)
denotes the RMS delay spread RMS normalized to the sampling rate 1/TC and is the amplitude
of the fading coefcient.
Assuming perfect power control and neglecting pathloss results in
L1
X
2 [k] = 1
(3.26)
k=0
The supported length L of the sampled impulse response should be chosen with respect to the
signal-to-noise ratio (SNR) ( E2sz ), at which the system is operating [28, p.8],
L1+
26
RMS
ln (SNR)
TC
(3.27)
Hence the components of the impulse response smaller than the SNR are not taken into account.
Generation of a channel impulse response when the power delay prole is known is achieved
assuming Rayleigh fading, which is described in next section.
2
2 ,
density function
p() =
p() =
22
e 2 , 0
2
1
,
2
(3.28)
(3.29)
where p() is the Rayleigh distribution, is the amplitude of the fading coefcient and the phase
uniformly distributed.
Each tap is hence modeled as Gaussian real and imaginary part with zero mean with a variance
according to the power delay prole. The goal is to model the time variation of the channel
properties too. One method to simulate this phenomenon is to use Jakes model [7]. However a
model developed by 3GPP, the Spatial Channel Model (SCM), described in the following Section
3.6.1, is used in this thesis in order to get as close to a true time varying channel as possible.
3.6.1
3GPP has developed a Spatial Channel Model (SCM) [2] for multiple input multiple output
(MIMO) systems in order to specify parameters intended for the three most common cellular
environments: suburban macrocells, urban macrocells, and urban microcells.
Since the goal of this project is channel estimation, pathloss is neglected.
The overall difference between the models are the amount and size of the scatterers. The macrocell
environments assume height of NodeB well above the height of scatterers, while the scatterers and
BS are at same height for the microcell scenario. From a UEs point of view the delay spread is
one of the main parameters affected by the change in channel model as shown in Table 3.1. SCM
has been implemented in Matlab [21] and can be used to generate channel fading coefcients,
depending on several parameters such as the speed of UE, number of antennas and their properties
27
Channel Model
0.251 s
0.65 s
0.17 s
Table 3.1: Root mean square delay spread for different channel models according to [2].
Figure 3.2: Block diagram of implementation of the SCME channel model with the transmitter and receiver.
transmitted signal is affected by the multipath channel generated by SCME and additive white
Gaussian noise (AWGN) is added afterwards. In Section 3.3.2 it was concluded that the channel
was time-invariant within one OFDM symbol duration. Hence in this thesis the SCME channel
model is congured such that channel impulse responses are generated for each OFDM symbol
and convolved with the symbol.
28
3.7.1
AWGN
Since the subcarriers of the transmitted signal are normalized, the signal-to-noise ratio is controlled
by adjusting the noise variance z2 . The average energy of one OFDM symbol is calculated as
following,
ES =
NBW
.
NIFFT
(3.30)
3.7.2
ES
NBW 1
=
.
EN
NIFFT z2
(3.31)
The impulse response generated by the SCME model is based on 6 paths at different delays, which
depend on the chosen channel scenario. The resolution of the delays is of
1
163.84 MHz
= 0.0163 s.
For a given sampling rate Tc the impulse response is assumed lowpass ltered with cutoff at
fcut =
1
2Tc
The lowpass ltering yields a convolution with a sinc in time domain, hence the generated impulse
response is ltered with a sinc, with width Tc .
In Table 2.1 different sampling frequencies are listed for the LTE downlink. For Tc =
1
3.84 MHz ,
the essential support of taps is chosen as L = 12 resulting in a power delay prole depicted in
Figure 3.3. The power delay prole is calculated over 2000 different channels generated by the
SCME model. The simulations will be carried out in the SNR range 0 to 20 dB and the minimum
noise level is also depicted on the gure. Hence components of the channel impulse response that
are smaller than the minimum noise variance are not taken into account.
0
10
Power of component k
Noise variance z2 at SNR=20 dB
|h[k]|2
10
10
6
Component k
10
29
3.8 Summary
In order to do simulations as close to the reality as possible, it is important to have a good channel
model. Mobile channel models with time invariant and time variant behavior have been investigated. The chosen channel model is the SCME (Spatial Channel Model Extended), which generates channel coefcients based on the 3GPP channel model specications. The channel model
supports a typical urban area scenario as well as mobility of the UE. Moreover the selection of
the impulse response length and adapting the impulse response to a sampling frequency Tc has
been described in order perform simulation of channel estimation methods with the LTE downlink
structure.
30
mation
This chapter describes different channel estimation techniques to be used for the time-invariant
downlink case.
By focusing on time-invariant channels, only the frequency-domain needs to be considered. Hence
three different channel estimators are applied in the frequency-domain and compared. Since the
estimation is aimed for the UE, the complexity should be minimized.
(4.1)
In order to ease the calculation of the channel estimation, the signal model is simplied in the
following section.
4.1.1
The channel is estimated for each OFDM symbol and the general focus of the channel estimation
will be on only one OFDM symbol. For the sake of simplicity the index m is dropped and equation
(4.1) is rewritten as
y = XF h + z,
(4.2)
where y CNIFFT is the received OFDM symbol, X CNIFFT ,NIFFT is a diagonal matrix with data,
reference symbols or zeros while h CNIFFT is the channel impulse response and z CNIFFT is
the noise which is assumed to be white Gaussian. F CNIFFT NIFFT is the same DFT-matrix as
described in (A.13),
F =
f1,1
..
.
...
..
.
f1,NIFFT
..
.
(4.3)
The channel estimation will be based on transmitted reference symbols, which are scattered as
described in Section 2.4.2, i.e. an interpolation has to be performed in order to obtain the channel
estimates for the subcarriers with data symbols. Before the rst estimation method is introduced,
prior knowledge about the channel is used to reduce the complexity of the estimator.
Based on the knowledge of the channel model and of the reference symbols two following assumptions can be made [24]:
- If the channel h has the maximum delay at tap L 1, only the rst L columns of F can be
considered, since the rest is multiplied by zero. h CL denotes the rst L coefcients of
h.
- The transmitted reference symbols are scattered (see Figure 2.4), hence only rows corresponding to the position of these symbols need to be considered for the diagonal matrix
X.
Equation (4.2) is rewritten as
yr = Xr Tr h + zr ,
where Xr = diag (xr (1) . . . xr (Nr )) and
..
..
..
Tr =
.
.
.
ftr (Nr ),1 . . . ftr (Nr ),L
(4.4)
(4.5)
Nr denotes the number of reference symbols for one OFDM symbol, while xr (i) and tr (i) denote
the ith reference symbol in the frequency-domain and the index of the subcarrier carring the
symbol respectively. zr CNr is the truncated noise and is still white Gaussian.
(4.6)
where gLS CNr can only be estimated over the subcarriers carrying reference symbols. Hence
this result has to be interpolated over the full frequency range in order to give an estimate for
subcarriers with data symbols.
The interpolation can be performed in frequency domain or in time domain. For the latter case the
interpolation is carried out by only taking the rst L channel taps into account and setting all other
taps to zero. After estimating the L-tap channel, the channel is transformed back into frequency
domain. This approach is investigated by applying linear minimum mean square error (LMMSE)
estimator in Section 4.3 and least-squares (LS) estimator in Section 4.4.
32
y
h = Rhyr Ry1
r yr r
(4.7)
R yr yr
= E yr yr H
= Xr Tr E hhH XrH Tr H + E zr zr H + Xr Tr hzr H + nr hH TrH Xr H
(4.8)
(4.9)
In order to simplify calculations it is assumed that the channel coefcients are independent and
the energy for each tap is 1/L, i.e. the power delay prole is uniform Rh h = E h hH = L1 I.
Rh yr
i
h
H
= E h yr H = E h h TrH XrH + zr hH TrH XrH
= Rh h TrH XrH
1 H H
=
T X .
L r r
(4.10)
(4.11)
(4.12)
..
..
..
.
Td =
.
.
.
(4.13)
Nd denotes the number of data symbols for one OFDM symbol, while xd (i) and td (i) denote the
ith data symbol and the index of the subband carring the symbol respectively. zd CNd is the
truncated Gaussian noise.
The calculated value for h is transformed into frequency domain for the subcarriers containing
data symbols,
g = Td
h Td h .
(4.14)
The frequency response of the channel is now interpolated using LMMSE estimation with low
complexity and the estimate is passed to the equalizer. It should be noted that for the case with
the true autocorrelation Rh h and not uniform as assumed this method is equal to the Wiener
interpolation [14].
33
4.4 Downsampling
Section 4.2 gives an estimate for the LS frequency response of the channel. Now a LS estimate
in the time domain is calculated. By modifying (4.2) it is transformed into the time domain. To
understand the idea behind this method, let us consider the received signal r in time for one OFDM
symbol.
r = F H XFL h + zt ,
(4.15)
where h CL is the L-tap channel impulse response and FL CNIFFT L is the Fourier matrix
that gives the frequency domain representation over NIFFT subcarriers of the channel, while X is
the transmitted symbols.
F H CNIFFT NIFFT transforms the attenuated signal in frequency domain into time domain, where
r is the time domain representation of y from (4.2) and zt is the complex Gaussian noise.
4.4.1
(4.16)
S = F H XFL ,
(4.17)
where
and the diagonal matrix X containing the complex symbols modulated over the subcarriers can
be expressed as:
X = Ad + Ar ,
(4.18)
where Ad CNIFFT NIFFT and Ar CNIFFT NIFFT are diagonal matrices containing non-zero elements in the positions of the transmitted data and of the transmitted reference symbols respectively.
Since the transmitted symbols are unknown, an approximation of the matrix S is made such that
only transmitted reference symbols are taken into account:
b = F H Ar FL .
S
(4.19)
(4.20)
hLS = FLH AH
r Ar FL
1
H H
FLH AH
r F r = FL Ar Ar FL
1
FLH AH
r y
From a computational point of view the LS estimator is simple since the matrix
FLH AH
r Ar FL
34
1
FLH AH
r
(4.21)
4.4. DOWNSAMPLING
is constant. In the LTE application, the matrix inversion can be computed once and used regardless
of the varying channel statistics. This of course is based on the assumption that the positions
of the reference symbols do not change. However, an ill-conditioned problem caused by the
matrix inversion occurs in the implementation of the LS estimator. The problem can be solved by
regularizing the eigenvalues of the matrix to be inverted by adding a small constant term to the
diagonal [6] or by downsampling.
4.4.2
The ill-conditioning problem stems from the structure of the subcarriers in the LTE downlink.
Only the center of the frequency spectrum is utilized, while the rest is set to zero.
As an example the case of NIFFT = 2048 from Table 2.1 is considered. The number of occupied
subcarriers is only 1200. Hence, while the sampling frequency is 30.72 MHz (NIFFT fspace ),
the occupied bandwidth is only 18 MHz (NBW fspace ). In LTE the occupied bandwidth is ap-
proximately 2/3 of the whole bandwidth. The goal is to get the occupied bandwidth close to the
sampling frequency which can be accomplished by downsampling to 2/3 of the sampling frequency. In practice, the channel h is not estimated in all the L taps but only in 2 out of 3 taps
hereby obtaining the average downsampling factor 2/3, as described in (4.23).
gDS
f1,1
f1,2
f1,3
= FL h =
f
4,1 f4,2 f4,3
.
..
..
..
.
.
= F DS hDS
f1,1
f1,2
...
f1,L
...
f2,L
...
f3,L
...
..
.
f4,L
..
.
. . . fN,L
f1,4
...
f1,L
h0
h1
h3
h4
..
.
hL
h0
f2,L
h1
. . . f3,L h3
. . . f4,L
h4
..
.. ..
.
. .
. . . fN,L
hL1
...
(4.22)
(4.23)
By removing each third column in FL resulting in F DS CNIFFT LDS the calculations are simpli-
r = F H Ar F DS hDS + zt ,
(4.24)
35
F DS,H Ar H Ar F DS
1
F DS,H Ar H y.
(4.25)
The channel impulse response hDS is transformed into frequency domain by multiplying with
F DS , such that
g DS = FdDS hDS = FdDS F DS,H Ar H Ar F DS
1
F DS,H Ar H y,
(4.26)
where FdDS is the truncated version of F DS , only containing rows coresponding to positions of
the data symbols. Utilizing knowledge of the transmitted reference symbols the complexity of
equations (4.25) and (4.26) can be reduced even further.
4.4.3
In [6] the downsampling to 2/3 of the sampling frequency is performed by estimating h only in
2 out of 3 taps.
In order to explain the purpose of this step, the downsampling procedure is investigated. Figure 4.1
shows an equivalent representation the downsampling of the non-integer factor 2/3 is performed
[17]. h is upsampled by a factor of two as follows,
Figure 4.1: System for changing the sampling rate by a non-integer factor [17].
(4.27)
The frequency response of the upsampled impulse response is illustrated on Figure 4.2(b). The
frequency response of the channel is only illustrated as rectangular as an example. The one-sided
bandwidth of the channel is assumed to be at most 2/3 of half the sampling frequency,
2
fBW = fs ,
3
where fs =
(4.28)
fs
2.
After 2x upsampling to sampling frequency fo = 2fs the maximum one sided bandwidth of the
channel is now at most 1/3 of half the new sampling frequency,
1
= fo .
fBW
3
36
(4.29)
4.4. DOWNSAMPLING
where fo =
fo
2.
In order to avoid aliasing when downsampling with factor three, a lowpass lter is introduced as
in Figure 4.2(c). The method proposed in [6] is equivalent to the following lowpass lter:
hlp = [1, 1] ,
(4.30)
ho = [h[0], h[0], h[1], h[1], h[2], h[2], h[3], h[3], . . . , h[L 1], h[L 1]] .
(4.31)
Ideally the lowpass lter should have cutoff at 1/3fo [17] but using hlp results in a sinc-function
in frequency domain with an one-sided bandwidth of fo which is depicted in Figure 4.2(c). The
h[0], h[0], h[1], h[1], h[2], h[2], h[3], h[3], . . . , h[L 1], h[L 1]
#
(4.32)
where it is assumed that h[L 1] is also an element to be selected in the downsampling. Equation
(4.32) is exactly the result described in Section 4.4.2 equation (4.23). The downsampled sampled
37
4.4.4
Complexity Reduction
From the structure of the reference signals described in Section 2.4.2 it is known that the norm is
1. Hence the term Ar H Ar in (4.25) consists of ones at reference positions and zeros otherwise.
The matrix inversion is rewritten according to this observation,
F DS,H Ar H Ar F DS
1
1
Fr DS,H ,
(4.33)
where FrDS is the truncated version of FLDS only containing rows corresponding to reference symbol positions. Equation (4.26) with reduced complexity is now expressed as
gdDS = FdDS hDS = FdDS Fr DS,H FrDS
1
Fr DS,H Xr H yr ,
(4.34)
with XrH and yr only containing values corresponding to reference positions as described in Section 4.1.1.
(4.35)
where gLS CNr is the least square channel frequency response, estimated using transmitted
reference symbols as in (4.6). To be able to calculate the autocorrelation RggLS between the full
range, g CNIFFT , and least squares frequency response, the vector gLS is set as a function of g
yr = Xr Tr
1
FL H g + zr
NIFFT
(4.36)
channel taps and N is the DFT-size. The frequencies corresponding to the subcarriers of the
reference symbols are then extracted by Tr and multiplied with the transmitted reference symbols.
gLS can now be expressed as
gLS = Tr
38
1
FL H g + Xr 1 zr .
NIFFT
(4.37)
= E ggLS H
1
= E g gH
FL Tr H + zr H X 1,H
NIFFT
1
=
FL Tr H
NIFFT
As in (4.9) each tap in the channel impulse response is assumed to have energy L1 , hence
1
E gg H = FL Rh h FLH = FL FLH .
L
The same approach is used to calculate RgLS gLS .
RgLS gLS = E gLS gLS H
1
1
H
H
1
H
1,H
H
FL gg
FL Tr + Xr zr zr Xr
= E Tr
NIFFT
NIFFT
h
1 i
1
=
Tr FL H FL Tr H + E z 2 Xr Xr H
NIFFT
L
Tr Tr H + z 2 I,
=
NIFFT
(4.38)
(4.39)
(4.40)
(4.41)
(4.42)
The average energy of the transmitted symbols are normalized to one, hence E Xr Xr H = IL .
In order to reduce the rank of the LMMSE estimate, (4.35) is factorized into [8],
gLMMSE = RggLS Rg0.5
Rg0.5
g .
LS gLS
LS gLS LS
(4.43)
(4.44)
"
Dp 0
0
Q2 H Rg0.5
g
LS gLS LS
g
= Q1 D Q2 Rg0.5
LS gLS LS
H
(4.45)
where Dp is the upper p p left corner of D, i.e. only the singular vectors associated to the
p largest singular values are kept and the rest is set to zero. Hereby noise can be reduced by
neglecting subspaces with low energy. The rank p is set as the estimated number of channel taps
(p = L) [8]. It is also possible to further simplify (4.45) according to [18] in order to reduce the
complexity. Equation (4.45) however is implemented in Matlab in order to compare this method
with others.
xed signal to noise ratio however we can precompute most of the calculation and hereby reduce
the calculation complexity. In this case the LMMSE CIR estimator requires step 3 and 4, while
the reduced rank LMMSE estimator requires step 5. This yields a low-complexity approach from
a calculation point of view to estimate the channel frequency response.
4.8 Summary
Estimation methods for calculation of the frequency response have been presented for the LTE
downlink in the time-invariant channel. The performed estimation is based on transmitted reference symbols. The presented algorithms are LMMSE CIR estimator, downsampled CIR estimator,
reduced rank LMMSE estimator and linear interpolation. The LMMSE CIR and downsampled
CIR present a way to estimate the channel impulse response based on assumptions of the number
of taps L. The downsampled CIR furthermore makes use of the fact that the LTE downlink only
occupies 2/3 of the transmitted bandwidth. The complexity of the algorithms have been reviewed
and it is shown that the reduced rank LMMSE has the highest complexity, followed by LMMSE
CIR and downsampled CIR. It is also shown that with a xed SNR we can reduce the calculation
complexity for the LMMSE CIR estimator and the reduced rank LMMSE estimator. The mobile
channel is not time-invariant hence a method has to be found in order to also perform estimation
in a time-varying channel. The next chapter introduces a method to cope with such channels.
40
4.8. SUMMARY
Estimator
steps
Calculation
Complexity
Memory
LMMSE CIR
Ry1
r yr :
O(Nr3 )
Downsampled CIR
O(LNr )
gest = fft(h):
O(NIFFT log(NIFFT ))
= Rg0.5
Rg0.5
:
Rg1
LS gLS
LS gLS
LS gLS
O(Nr3 )
:
= RggLS Rg0.5
LS gLS
svd():
O(NBW Nr )
= Q1 D Q2 H :
:
= Rg0.5
LS gLS
gest = gLS :
O(NBW Nr )
hDS
O(LDS Nr )
O(NBW LDS )
Linear Interpolation
O(Nr 1)
= yr :
O(NBW
Nr2 )
=
= yr :
h
Rh yr Ry1
r yr :
O(NBW Nr2 )
Rh yr CLNr
Tr Tr H CNr Nr
RggLS CNBW Nr
O(NBW L Nr )
O(NBW Nr2 )
1
= FrDS,H FrDS
FrDS,H Xr H CLNr
FTDS CNBW LDS
-
Table 4.1: Complexity of frequency response calculation for the different methods described in this chapter.
The required memory for complex elements is calculated as the product of the matrix dimensions. For the
LMMSE CIR estimator and reduced rank LMMSE estimator we only have to perform the calculation after
the horizontal line for a fixed signal to noise ratio, since all the other computations can be precomputed.
41
42
tion
The mobility of the UE and the surrounding objects leads to a time-varying channel which needs
to be tracked. As described in Section 3.3.1 this results in a Doppler shift. In order to estimate the
time-varying channel, the discrete prolate spheroidal (DPS) basis expansion is utilized. Firstly, the
general idea behind time-varying channel estimation is introduced before describing the utilization
of the DPS basis expansion in detail.
5.1.1
In order to track the channel frequency response, the subcarriers with reference symbols are used
to nd estimates of the channel for the subcarriers at OFDM symbols without reference symbols.
This can be illustrated by Figure 5.1. Firstly, the channel is estimated in the time direction of all
subcarriers with reference symbols using DPS sequences. This method allows a channel estimate
to be found for each third subcarrier at each OFDM symbol. Secondly, the frequency response
for each OFDM symbol is found by interpolating in the frequency-domain between the known
estimates. Let us denote the number of reference symbols for subcarriers and OFDM symbols
carrying them as Nr and Mr respectively. The notation M is used in order to emphasize that the
reference symbols in time-domain are considered.
Figure 5.1: Illustration of the estimation procedure. The horizontal axis is the channel frequency-response
also referred to as frequency-domain, while the vertical axis corresponds to the time-varying behavior of
the channel. The vertical dimension is referred to as time-dimension. The positions of reference symbols is
according to Chapter 2 Section 2.4.2.
Interpolation
Three interpolation methods are utilized.
- Linear interpolation between estimated subcarrier coefcients.
- Interpolation using the reduced rank LMMSE estimator.
- Interpolation using the DPS basis expansion.
The rst two estimators are already known from Section 3.2.1 and 4.5 respectively. DPS sequences
can be applied in the frequency dimension as well, which will be described in this chapter.
5.1.2
Instead of tracking the channel coefcients for subcarriers with reference symbols, the taps of
the channel impulse response can be tracked, i.e we estimate the channel impulse response in the
frequency-domain for OFDM symbols with reference symbols and then we track each tap in the
time-domain.
Sections 4.3 and 4.4.2 presented two different approaches to estimate the channel impulse response. The task is now to rstly estimate a L-tap channel impulse response for each OFDM
symbol carrying reference symbols. Hereafter each tap is tracked using DPS sequences instead
of the subcarriers with channel estimates. In order to estimate the channel frequency response for
each OFDM symbol, a FFT is performed on the estimated impulse response.
44
Common for all estimation methods is the use of DPS sequences, which will be explained in the
following section.
2 ,
|u[m]|
with 0 1.
(5.1)
(5.2)
Dmax
U () =
u[m]ej2m .
(5.3)
The solutions for the optimization problems (5.1) (5.2) and (5.3) are the discrete prolate spheroidal
sequences. The sequences ui [, Dmax , M ] are dened as the real solution of
M
1
X
=0
sin(2Dmax ( m))
ui [, W, M ] = i (Dmax , M )ui [m, W, M ]
( m)
(5.4)
for i {0, . . . , M 1} = I and m Z. In the sequel a system with xed parameters (W, M ) is
The DPS sequence ui [m] is a unique sequence which is exactly band-limited and having a high
energy concentration within the time interval M . The sequences are doubly orthonormal on the
innite set {, . . . , } = Z as well as the nite set {0, . . . , M 1},
M
1
X
ui [m]uj [m] = i
m=0
ui [m]uj [m] = ij ,
(5.5)
where i, j {0, . . . , M 1}. The eigenvalues i of the sequences ui [m], expressed by (5.1)
(5.6)
For the purpose of channel estimation, only the index set m {0, . . . , M 1} is of concern and
these limited sequences are denoted as Slepian sequences. The vector ui RM is now dened
containing elements ui [m] for m {0, . . . , M 1} and equation (5.4) is written as:
Cui = i ui ,
(5.7)
sin(2Dmax (i ))
,
(i )
(5.8)
D1
X
ui [m]i ,
(5.9)
i=0
where i is the weighting coefcients of the Slepian sequences and m {0, . . . , M 1}. The
dimension D has the constraints,
D D M.
(5.10)
The choice of D affects the mean square error (MSE) dened as,
MSEM
M 1
o
1 X n
=
E |g[m] g[m]|2 .
M
(5.11)
m=0
5.2.1
In order to utilize a basis expansion for channel estimation, the received OFDM signal is investigated. The received signal y[m, q] for OFDM symbol at time index m for subcarrier q. For a xed
subcarrier, the index q is dropped resulting in
y[m] = g[m]d[m] + z[m],
(5.12)
where g[m] is the channel attenuation, d[m] the transmitted symbol and z[m] is white Gaussian
noise. Using a basis expansion (5.12) can be written as,
!
D1
X
ui [m]i d[m] + z[m].
y[m]
i=0
46
(5.13)
(5.14)
M
1
X
g[m]ui =
m=0
M
1
X
(5.15)
m=0
g[m] = f [m] =
D1
X
ui [m]
i
i=0
(5.16)
where
f [m] =
and
ui [m]
..
.
uD1 [m]
i
..
.
D1
RD
(5.17)
CD .
(5.18)
It can be shown that the Slepian basis expansion is a reduced-rank channel estimation. The following section describes reduced-rank estimation in general.
5.2.2
Reduced-rank channel estimation is expressed in equation (5.9), which also can be written as [27],
g[m] g[m] =
D1
X
ui [m]
i = U
,
where
(5.19)
i=0
U = [u0 , . . . , uD1 ] ,
ui = [ui [0], ui [1], . . . , ui [M 1]]T ,
with
i {0, . . . , D 1} .
(5.20)
(5.21)
(5.22)
because of the orthonormality of the basis vectors. The reconstruction error per data block is
dened as [27],
z=
1
1
U H n
2 + 1 V H g 2 ,
||g
g | |2 =
M
M
M
(5.23)
47
= E {z}
D 2
1 n H 2 o
=
.
n +
E V h
M
M
(5.24)
When using Slepian sequences, the subspace dimension D must be chosen such that reconstruction
error is minimized.
5.2.3
In the LTE downlink the reference symbols are only located at some OFDM symbols m Pq ,
where Pq is the set of positions of reference symbols of subcarrier q as depicted in Figure 5.1. Let
us consider a xed subcarrier q and neglect the notation q. Equation (5.15) is dened for a signal
known for all m I. The orthogonality of the Slepian sequences is lost by only taking some
values of m. In order to correct the loss of orthogonality the instantaneous values the correlation
matrix G is calculated [26].
G=
mP
f [m]f H [m],
(5.25)
mP
where p[m] is the value of the transmitted reference symbol at time m. According to Section 2.4.2
the norm of the reference symbols are always one, |p[m]| = 1.
where
(5.26)
mP
= [
0 , . . . , D1 ]T .
(5.27)
Let us dene Ur as the truncated version of U with only rows corresponding to positions of
reference symbols. Using the notation from Section 5.2.2, (5.25) is rewritten as follows,
G = UrH Ur ,
(5.28)
= G1 UrH gmr ,
(5.29)
where gmr CMr is the vector containing the scattered estimates of the channel only.
The mean square error can be described by a square bias and a variance term [26],
M SEM = bias2M + varM ,
(5.30)
where bias2M depends on the actual number of basis functions while varM depends linearly on the
noise term, the number of reference symbols J, and the number of basis functions D:
48
varM z2
M
1
X
m=0
f H [m]G1 f [m] z2
D
.
J
(5.31)
vmax fC
Ts = 0.0185.
c0
(5.32)
(5.33)
(5.34)
(5.35)
There are two slots per subframe with each slot having 6 OFDM symbols (see Section 2.2). The
total number of OFDM symbols (block length) is
M = 2Nsub 6 = 48.
(5.36)
49
There are two OFDM symbols with reference symbols within one slot, but the OFDM symbols
do not have the reference symbols at same subcarrier positions. Hence only one reference OFDM
symbol represents one reference subcarrier for each slot.
Mr = 2Nsub = 8.
(5.37)
The Slepian sequences are generated in Matlab with the command dpss. For this example the inputs are as follows dpss(M ,Dmax M ,D). With D set to 4 as an example, the generated Slepian
sequences are depicted on Figure 5.2.
0.4
0.3
0.2
0.1
0
0.1
0.2
u0 [m]
u1 [m]
0.3
u2 [m]
u3 [m]
0.4
0
10
15
20
25
30
35
40
45
Figure 5.2: Slepian sequences for ui [m] for M = 48 and Dmax = 0.0185.
5.3.1
In Section 5.1.1 it was mentioned that Slepian Sequences could be applied in the frequency dimension as well. This is achieved as follows [25],
- For a xed OFDM symbol m, the frequency response g CNIFFT spans a subspace that has
dual properties to the Slepian subspace in the time-domain.
L1
NIFFT
Dmax .
The generated Slepian sequence here denoted as u
i CNIFFT for the frequency domain has the
dimension Df max NIFFT + 1 = L. For the time-domain the Doppler spectrum is symmetric
We denote the usage of Slepian sequences to estimate the frequency domain as 2x1D Slepian,
since Slepian sequences are applied in the frequency-domain as well as in the time-domain.
5.5.1
Interpolation in Frequency-Dimension
The interpolation in frequency is similar to Section 4.3 where LMMSE estimation was used. Now
the true autocorrelation of the channel is used, instead of assuming a uniform power delay prole.
The interpolated frequency response, gI , is written as,
y
gI = Rgyr Ry1
r yr r
= F Rh h TrH XrH Xr Tr Rh h TrH XrH + z 2 INr
(5.38)
1
yr
(5.39)
5.5.2
Interpolation in Time-Dimension
The interpolation in the time direction requires information about the correlation, which depends
on the Doppler shift. Let y[m, q] denote a received value at time m for subcarrier q. For a xed
subcarrier, index q is dropped.
The signal y[m] is only known at positions with reference symbols (m P). ymr CP denotes
the vector containing only received reference symbols, while ym CM contains the values of
y[m]. Index m is added such that this vector is not confused with the notation yr from Section
5.5.1. The channel frequency response is sought, hence it is estimated for all positions with reference symbols (m P):
gmr = Xmr ymr ,
(5.40)
where the matrix Xmr CMr is the diagonal matrix containing the transmitted reference symbols
on the time-dimension for a xed subcarrier q and Mr is the number of reference symbols. The
interpolated estimate in the time-dimension using Wiener interpolation is [14]:
gm = Rgm gmr Rg1
gmr ,
mr gmr
(5.41)
where
gm CM . The next task is to derive the values of the correlation matrices.
Rggmr C M,P is the cross-correlation between g and gmr . The following example shows how the
cross-correlation is found.
5.5.2.1
For the scalar case g[m] with reference symbols for all values of m, the interpolated value is found
as,
1
gm .
g[m] = rgm gm [m]T Rg1
m gm
(5.42)
(5.43)
The index r is dropped for indicating all reference symbols. The correlation vector rgm gm [m] can
be written as,
rgm gm [1]
rgm gm [M 2]
rgm gm [0]
=
..
.
..
.
...
rgm gm [0]
rgm gm [M + 1] rgm gm [M ] . . .
(5.44)
(5.45)
(5.46)
The estimated values for the channel is found on each third OFDM symbol since an interpolation
in frequency direction was applied rst, hence the crosscorrelation rgm gmr yields,
rgm gmr [m] = [rgm gm [m], rgm gm [m + 3], rgm gm [m + 6], . . . , rgm gm [m + 46]] .
(5.47)
Using this knowledge, Rgm gmr is found knowing Rgm gm . The rows of Rgm gmr are the same as
for Rgm gm , however, only the columns of Rgm gm corresponding to the positions of the reference
symbols taken into account.
The autocorrelation matrix is calculated next. It can be shown that
Rgmr gmr = Rgmr gmr + IM r z2 ,
(5.48)
where Rgmr gmr is found using the knowledge acquired from (5.44) and using it for the scattered
case,
Rgmr gmr = toeplitz {[rgm gm [0], rgm gm [3], rgm gm [6], . . . , rgm gm [3(Mr 1)]]}
(5.49)
The question is now how to estimate the correlation vector rgm gmr [m]. An approximated analytical
expression for the autocorrelation has been found in Section 3.4. However the autocorrelation for
the SCME model is found through simulation, in order to obtain a better estimate.
5.6.1
The Wiener interpolation for all reference symbols is expressed in (5.42). If covariance matrix
1
Rg1
m gm = Rgm gm that is if the noise is neglected and the optimum dimension is D, the estimated
D1
X
ui i ,
(5.50)
i=0
53
D[1,...,M ]
M 1
1 X
D 2
i +
M
M z
i=D
"
(5.51)
Rg [ m]ui [] = i ui [m],
m Z.
(5.52)
The sequences ui for m IM form orthogonal sets. Based on different assumptions on the
5.6.2
The utilization of Slepian sequences in based on the assumption of a flat Doppler spectrum
S(, W). In order to relate it to a general reduced rank estimator, let us consider the sequence
ui (m, W). The index W is included to state that the Slepian sequences depends on the Doppler
spectrum, which is band-limited in the region W. In (5.7) and (5.8) it was shown how the Slepian
sequences were found. Comparing this information to the covariance matrix of a given Doppler
spectrum expressed in (3.23) in Section 3.4 yields,
C[k, W] = |W|Rd [k, W]
(5.53)
We now express how the basis vectors of the Slepian sequences depend on Rd .
ui (W) =
M 1
|W| X
Rd [ m] ui [, W]
i (W)
=0
|W|
rd [m, W]H uW ,
i (W)
(5.54)
where the knowledge of C[k, W] = |W|Rd [k, W] is first used and then Rd [k, W] = Rd [k, W]
is applied [27]. Equation (5.50) and (5.17) associated with (5.54) yields
f [m, W]T
(5.55)
1
|W| diag (0 , . . . , D1 )
and
U (W) = [u0 (W), . . . , uD1 (W)]. By inserting the information from equation (5.55) into the
sequence estimation (5.16), equation (5.56) is obtained. It should be noticed that in (5.16) the
index W is dropped, but it is used here.
g = f T [m, W]
H
(5.56)
1
= rd [m, W] U (W)
(W)U (W) y,
(5.57)
which is exactly the same as (5.50) with reduced rank estimator for a constant Doppler spectrum,
Rg [k] = Rd [k], for k Z. Hence both estimators are using the same subspace spanned by
5.7. COMPLEXITY
5.7 Complexity
The complexity of the time-varying channel estimation per data block of is divided into two
parts, the first part is the estimation in the time-dimension and the second part is the frequencydimension. This section considers the complexity in the time-dimension, since Section 4.7 already
describes the complexity for estimation in the frequency-dimension. The complexity of Slepian
sequences depends on whether the estimation is performed for the subcarriers with channel estimates or the impulse response. In general the complexity in the time-dimension for one subcarrier
or tap can be expressed as in Table 5.1. The complexity of the projection depends on the number
Steps
Calculation
Complexity
Memory
mr :
= UrH g
= G
:
D Mr
U H CM D
g = U
D2
DM
G CDD
5.7.1
The complexity of channel estimation by estimating the subcarriers rstly and then interpolating
as described in Section 5.1.1 depends on which interpolation method is utilized and on the number
of subcarriers to estimate.
The number of subcarriers with reference symbols is NR = 2Nr , since scattered reference
symbols are applied, see also Figure 5.1. The calculation described in Table 5.1 has to be performed 2Nr times. After the rst estimation in the time-domain there are 2Nr estimates of the
frequency response for each OFDM symbol. Hence Nr has to be replaced with NR for reduced
rank LMMSE estimator and linear interpolation in Table 4.7 and the stored matrices have to include the new positions of estimates as well in order to take them into account in the calculation.
5.7.1.1 Slepian Sequences in Frequency Dimension
The complexity of applying Slepian sequences in the frequency-dimension is not expressed in
Table 4.7, hence it is calculated in Table 5.2.
5.7.2
Instead of estimating in the time-domain rstly, we estimate the channel impulse response in the
frequency-domain for each OFDM symbol with reference symbols and then perform estimation
in the time-domain. The number of OFDM symbols for which the taps are calculated is MR =
2Mr = 16, see also Figure 5.1 and (5.37). The complexity of the calculation for each OFDM
symbol is described in Table 4.7 for LMMSE CIR and downsampled CIR. Both impulse responses
can be transformed into the frequency-domain and the channel estimates for each subcarrier can
55
Steps
Calculation
Complexity
Memory
r :
f = UrH g
Df NR
UfH CNBW Df
:
G1
f =
g = Uf
Df2
Df NBW
G1 CDf Df
Table 5.2: Complexity of frequency response estimation using the Slepian sequences.
Steps
Calculation
Complexity
Memory
mr :
= UrH g
2DMr
= G1
D2
U H CM D
g = U
DM
G1 CDD
Table 5.3: Complexity of tracking the CIR using the Slepian sequences.
be tracked. However it is chosen to track the taps of the impulse responses in the time-varying
channel, because the number of taps is lower than the number of subcarriers. The complexity
of applying the Slepian sequences is calculated using the same steps as in Table 5.1. Instead of
considering one subcarrier we consider one tap. Table 5.3 lists the calculations for tracking one
tap from the impulse response in the time-domain. Since we need to track all taps, the calculations
in Table 5.3 need to be carried out L times and the found impulse response is transformed into the
frequency domain M times.
5.7.3
Table 5.4 lists the calculation for the Wiener interpolation, which is the estimator that yields the
lowest MSE, but has also the highest complexity, specially since it requires operation of an order
of 3, i.e O(MR3 ) for step 1 . Hence the usage of Slepian sequences in the time-domain offers a
low-complexity approach for estimation of the time-varying channel. By applying the Slepian sequences in time-domain for channel estimation and the different low-complexity estimators mentioned in Chapter 4 results in a low-complexity approach for the time-variant channel estimation.
56
5.8. SUMMARY
Steps
Calculation
Complexity
Memory
Rg1
mr :
mr g
O(MR3 )
gmr :
Rg1
mr
mr g
gm = Rgm gmr :
O(MR2 )
2
MM
R
5.8 Summary
In this chapter methods to estimate the time-varying channel for the LTE downlink were presented.
The channel is estimated using Slepian sequences in the time-dimension, while using LMMSE
CIR, downsampled CIR, reduced rank LMMSE, linear interpolation and Slepian sequences as
estimation methods in frequency-domain. Use of Slepian sequences in the time-domain corresponds to a reduced rank Wiener interpolation with a at Doppler spectrum. Furthermore the
Slepian sequences yields a low-complexity approach to estimate the time-varying channel in the
time-domain. In order to evaluate the performance of the presented methods, they are compared
to 2x1D Wiener interpolation with autocorrelation functions matching the channel. Combining
the low-complexity channel estimation approach in the time-domain introduced by the Slepian
sequences with the different estimators in the frequency-domain, we achieve a low-complexity alternative to the 2x1D Wiener interpolation. Beside the channel estimation it is also of importance
to predict the mobile channel. The next chapter considers how this can be achieved by expanding
the same methods utilized for channel estimation.
57
58
Channel Prediction
In Chapter 5 the channel estimation problem was considered, now channel prediction is investigated. The NodeB schedules resource blocks to the UEs for the each subframe. The scheduling is
based on the state of the channel, reported by the UE in a frequency division duplex (FDD) system.
However for mobile users at vehicular speed the channel state information gets outdated, hence
the channel must be predicted for a proper scheduling [20]. This chapter introduces a method to
predict the channel in UMTS LTE using Slepian sequences developed by [27].
denote the prediction as minimum-energy prediction since the Slepian sequences are concentrated
in the interval m IM . The used Slepian sequences for channel estimation purpose were dened
for m IM , hence the rst task is to obtain Slepian sequences at times > M 1.
6.1.1
By evaluating (5.4) in Section 5.2, basis vectors ui were calculated for a block of nite length, i.e.
ui [m] for m IM . The sequence ui [m] can be continued for m over Z in a minimum energy
(ME) band-limited sense by evaluating (5.4) on the right hand side, as expressed in (6.1) [27].
ui [, W, M ] =
M
1
X
m=0
1
sin(2Dmax (m ))
ui [m, W, M ]
.
(m )
i (W, M )
(6.1)
6.1.2
Prediction
given as
g[] =
D1
X
ui [l]
i ,
(6.2)
i=0
where ui [] is the extended basis sequence at time indices M .Using the same methods as
for channel estimation (Chapter 5), the channel can be predicted with the new knowledge from
(6.3)
where rgm gmr [m] is found the same way as in (5.47). The Wiener predictor can be approximated
by a reduced rank predictor where subspaces with small eigenvalues are truncated [27].
6.4 Summary
Using Slepian sequences it is possible to perform channel prediction such that the information
about the channel is not outdated at NodeB. Furthermore the use of Slepian sequences for prediction is similar to the reduced rank Wiener prediction.
60
Simulation Results
The presented estimation and prediction methods are now simulated using Matlab. The written
Matlab les can be found on the companion CD-ROM. The LTE downlink signal structure and
SCME model offer several parameters to adjust. It was chosen to work with the downlink transmission bandwidths of 2.5 and 20 MHz, since the rst use the same sampling frequency as the
current UMTS 3G system while the latter is the maximum transmission bandwidth.
7.1.1
The SCME model takes several inputs which are described in [21]. A channel has to be generated for each OFDM symbol, which has a duration of Ts as described in Section 3.3.2. [21] does
not give a detailed description of how to congure the model according the mentioned problem.
Therefore the solution to this is given in this section.
In light of line 531 in the Matlab le scm.m within the SCME package, the channel impulse response is generated for each OFDM symbol by conguring the parameter scmpar.SampleDensity
as follows,
SlotDuration=0.5e-3; %0.5 ms
N.OFDM_symbol_pr_slot=6; % 6 symbols pr. slot
speed_of_light=2.99792458e8; %m/s
fcarrier=2e9; % 2GHz carrier frequency
wavelength=speed_of_light/fcarrier;
delta_t=SlotDuration/N.OFDM_symbol_pr_slot; %OFDM symbol duration
linkpar.MsVelocity=channel.vmax; % Velocity of UE is configured
%Configuring parameter to generate impulse response for each OFDM symbol.
scmpar.SampleDensity = wavelength / (linkpar.MsVelocity*2*delta_t);
The parameter scmpar.SampleDensity determines the delay between the generation of each
channel impulse response.
0.8
0.6
r[k]
0.4
0.2
0.2
0.4
0
10
20
30
40
50
lag [k]
60
70
80
90
100
Figure 7.1: Autocorrelation for the SCME channel model. The correlation is averaged over 4000 generated
channels as well as over all taps.
autocorrelation when the Wiener interpolator is used (see Section 5.5.2). The power delay prole
is found as an average of power delay proles from 2000 generated channels.
Figure 7.2 shows the power delay prole for 20 MHz transmission bandwidth. The chosen
number of taps for this transmission bandwidth is 27.
10
Power of tap k
Noise variance z2 at SNR=20 dB
|h[k]|2
10
10
10
15
20
25
Tap k
Figure 7.2: Power delay profile for 20 MHz downlink transmission bandwidth using fs = 30.72 M Hz.
7.4 Evaluation
We apply the same procedure to investigate the 2.5 MHz bandwidth as for the 20 MHz bandwidth.
The used parameters are listed in Table 7.1. The power delay prole for 2.5 MHz transmission
bandwidth is depicted on Figure 3.3. For each OFDM symbol, the estimated frequency response
Parameters
Values
Channel model
Urban macro
Number of simulations
400
Taps
12 (27)
Cyclic Prex
Long
Block size
Sampling frequency
48 OFDM symbols
3.84 MHz (30.72 MHz)
4 (5)
120 km/h
12 (27)
Table 7.1: Parameters used in simulation for 2.5 MHz and 20 MHz downlink transmission bandwidth. For
the latter case the changed parameters are specified in the round brackets.
of the channel is compared to the true frequency response. The true frequency response is found
as the Fourier transform of the channel impulse response. The performance of the estimators
is measured in the MSE (mean square error) sense between the true channel and the estimated
channel. The frequency response is compared in the range of the used subcarriers (see also Figure
2.2), since the estimates are used for these positions only.
7.4.1
The achieved mean square error for different SNR values are depicted on Figure 7.3 for 2.5 MHz
transmission bandwidth and on Figure 7.4 for the 20 MHz bandwidth. The simulations are always
conducted over 400 different channel iterations.
For the 2.5 MHz transmission bandwidth it is noticeable that the performance of LMMSE CIR
estimator is close to the Wiener Interpolator. The downsampled CIR estimator again performs as
well as LMMSE CIR for SNR up to 10 dB, hereafter the performance decreases. The reduced
performance is caused by the aliasing due to the downsampling as described in Section 4.4.3. The
64
7.4. EVALUATION
10
2x1D Slepian
Linear interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced rank LMMSE estimator/Slepian
2x1D Wiener interpolation
Average MSE
10
10
10
8
10
12
Average SNR [dB]
14
16
18
20
Figure 7.3: MSE performance with 2.5 MHz transmission bandwidth at different SNR.
2x1D Slepian
10
Linear Interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced Rank LMMSE estimator/Slepian
2x1D Wiener interpolation
Average MSE
10
10
10
10
8
10
12
Average SNR [dB]
14
16
18
20
Figure 7.4: MSE performance with 20 MHz transmission bandwidth at different SNR.
reduced rank LMMSE estimator performs worse followed by 2x1D Slepian estimator. The linear
interpolation has the worst performance of all estimators.
65
For the 20 MHz transmission bandwidth in general the mean square error is lower compared
to 2.5 MHz transmission bandwidth. It is again noticeable that the performance of LMMSE CIR
estimator is closest to the performance of the 2x1D Wiener interpolator. The downsampled CIR
estimator performs as well as LMMSE CIR for SNR up to 10 dB. For higher SNR its performance
decreases as for the 2.5 MHz case. Linear interpolation has the worst performance which differs
by an order of one magnitude from the rest of the estimators.
7.4.2
The estimators must also be able to perform at speeds below 120 km/h, hence the performance
of channel estimation is investigated for the different estimators at speeds below 120 km/h but
always with Slepian sequences adapted for 120 km/h. Figure 7.5 depicts the performance from 0
km/h to 120 km/h for the 2.5 MHz transmission bandwidth. Since the 2x1D Wiener interpolator
is used as a reference its autocorrelation is adapted for the true speed. The Wiener interpolator
1
10
2x1D Slepian
Linear interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced rank LMMSE estimator/Slepian
2x1D Wiener interpolation
Average MSE
10
20
40
60
Velocity km/h
80
100
120
Figure 7.5: MSE for channel estimation with 2.5 MHz transmission bandwidth at different speed.
has the best performance, since the Doppler autocorrelation is adapted for each given speed. The
estimation errors of all estimators using Slepian sequences adapted for 120 km/h have the same
MSE for the different speeds. The LMMSE CIR estimator yields a better performance than the
other low-complexity estimators, while the linear interpolator has the worst performance.
In order to see the performance when the Slepian sequences are designed for the specic speeds,
the LMMSE CIR estimator is also applied with the mentioned sequences. The performance of
this estimator is closest to the 2x1D Wiener interpolator. Figure 7.6 depicts the performance for
66
7.4. EVALUATION
10
Average MSE
10
10
2x1D Slepian
Linear interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced rank LMMSE estimator/Slepian
10
20
40
60
Velocity km/h
80
100
120
Figure 7.6: MSE for channel estimation with 20 MHz transmission bandwidth at different speeds.
the 20 MHz transmission bandwidth. In this case the MSE of all estimators are lower compared
to 2.5 MHz transmission bandwidth. The Wiener interpolator has the lowest MSE, while the
estimation errors of all estimators using Slepian sequences adapted for 120 km/h have the same
MSE for the different speeds. We again apply the LMMSE CIR estimator with the adapted Slepian
sequences for the different speeds. The performance of this estimator is closest to the 2x1D Wiener
interpolator.
7.4.3
Channel Prediction
As mentioned in Chapter 6 channel prediction is also necessary. The prediction horizon is chosen
as 2 subframes, i.e. 24 OFDM symbols and the number OFDM symbols over which the estimation
is performed is still M = 48. Figure 7.7 depicts the measured MSE for 2.5 MHz transmission
bandwidth. The channel is predicted for a UE with velocity 120 km/h. Since the channel is
predicted after m = 47, the MSE is lower at m < 48 because the channel is estimated for this
interval, while the MSE increases sharply after m = 47 because of the channel is now predicted. In
general all estimators, except the linear interpolation, have the same MSE for channel prediction,
however the downsampled CIR estimator has the performance closest to the Wiener predictor. The
MSE performance for 20 MHz bandwidth is depicted on Figure 7.8. In this case the individual
performance of the estimators are in general the same as for 2.5 MHz transmission bandwidth.
The downsampled CIR estimator has the same performance for prediction as the other estimators.
The resulting MSE increases sharply after m = 47. All predicted values have in general the same
67
2x1D Slepian
Linear interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced rank LMMSE estimator/Slepian
10
Average MSE
10
10
10
20
30
40
50
OFDM symbol m
60
70
80
Figure 7.7: MSE for channel prediction with 2.5 MHz transmission bandwidth at 120 km/h. The prediction
starts after the thick vertical black line at m = 47. The 2x1D Wiener interpolator apply prediction for
m > 47 and estimation otherwise.
10
2x1D Slepian
Linear interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced rank LMMSE estimator/Slepian
Average MSE
10
10
10
10
10
20
30
40
50
OFDM symbol m
60
70
80
Figure 7.8: MSE for channel prediction with 20 MHz transmission bandwidth at 120 km/h. The channel
is predicted after m = 47.The 2x1D Wiener interpolator apply prediction for m > 47 and estimation
otherwise.
68
7.4. EVALUATION
MSE. Excecpt for the linear interpolation the prediction error has its maximum at m = 71. This
is the maximum horizon at which we can predict. This prediction horizon = 24 corresponds to
a covered distance in wavelength ,
d=
7
vmax Ts
= ,
c0 /fc
16
(7.1)
by the UE. Hence we can predict the channel for a movement of maximum
7
16
of the wavelength
= c0 /fc . It should be noted that we only calculate the distance at which we have the maximum
prediction error. For practical usage the distance, i.e. the prediction horizon may be reduced.
7.4.3.1 Channel Prediction at Different Speeds
As for the channel estimation in Section 7.4.2 we investigate the performance of channel prediction
at speeds below 120 km/h. The Slepian sequence used in time-domain is adapted for 120 km/h.
The prediction is performed for OFDM symbol m = 59, i.e. the horizon is one subframe (12
OFDM symbols). Figure 7.9 depicts the MSE for the 2.5 MHz transmission bandwidth. As for
10
Average MSE
10
10
2x1D Slepian
Linear interpolation/Slepian
LMMSE CIR estimator/Slepian
Downsampled CIR estimator/Slepian
Reduced rank LMMSE estimator/Slepian
LMMSE CIR estimator/Slepian w. known speed
2x1D Wiener Interpolation
10
20
40
60
Velocity km/h
80
100
120
Figure 7.9: MSE for channel prediction with 2.5 MHz at different speeds.
the estimation case the prediction errors have the MSE for the estimators with Slepian sequences
adapted for 120 km/h through out the range of chosen speeds. The downsampled CIR estimator
has the lowest MSE compared to the other low-complexity estimators. Figure 7.10 depicts the
same scenario for the 20 MHz transmission bandwidth. For this transmission bandwidth it is not
possible to discriminate between the performance of the low-complexity estimators. Again the
prediction errors have the MSE for the estimators with Slepian sequences adapted for 120 km/h.
69
10
Average MSE
10
10
2x1D Slepian
Linear interpolation/Slepian
10
10
20
40
60
Velocity km/h
80
100
120
Figure 7.10: MSE for prediction with 20 MHz transmission bandwidth at different speeds.
However if the speed is known as for the LMMSE estimator using Slepian sequences adapted to
the speed then the achieved MSE is lower and closest to the Wiener predictor.
7.5 Summary
The performance of the presented methods for time-variant channel estimation has been simulated
with the SCME channel model. The autocorrelation of the Doppler spectrum from the SCME
model is found numerically. The different estimators have been evaluated at 2.5 MHz and 20 MHz
downlink transmission bandwidths. All estimators are compared to the 2x1D Wiener interpolation
which is used as reference since it consistently yields the lowest MSE. In general the LMMSE
CIR estimator using Slepian sequences shows the best MSE performance of all estimators, since
it is closest to the 2x1D Wiener interpolator. Furthermore it is shown that the prediction horizon
for 2.5 MHz as well as 20 MHz transmission bandwidths corresponds to a covered distance of
approximately one half of the wavelength at 120 km/h.
70
In order to ensure competitiveness for the next few years a long-term evolution of the 3GPP radioaccess technology is currently developed. In this thesis important properties of the physical layer
in UMTS-LTE downlink have been investigated based on specications in Release 7 and on working assumptions in Release 8.
OFDMA is utilized as multiple access scheme in the downlink. One of the most important aspects
in an OFDM system is a reliable and accurate channel estimation.
In order to perform realistic simulations, channel models with time-invariant and time-varying behavior have been investigated. The chosen channel model is the Spatial Channel Model Extended
(SCME), which generates channel coefcients based on 3GPP channel model specications. The
channel model supports a typical urban area scenario as well as mobility of the UE. The SCME
model is implemented with the LTE downlink structure in Matlab.
Different low-complexity estimation methods have been presented, all based on Slepian sequences.
The channel is estimated using Slepian sequences in the time-dimension, while using LMMSE
CIR estimator, downsampled CIR estimator, reduced rank LMMSE estimator, linear interpolation
and also Slepian sequences as estimation method in the frequency-domain.
The LMMSE CIR and downsampled CIR present a way to estimate the channel impulse response
based on assumptions of the number of taps L. The downsampled CIR furthermore makes use of
the fact that the LTE downlink only occupies 2/3 of the bandwidth.
The complexity of the algorithms have been reviewed and it is shown that the reduced rank
LMMSE has the highest complexity, followed by LMMSE CIR estimator and downsampled CIR
estimator respectively.
In order to investigate the performance of the presented methods, they are compared to 2x1D
Wiener interpolation with perfectly known autocorrelation functions. Using Slepian sequences
it is also possible to perform channel prediction in a minimum-energy (ME) band-limited sense
such that the information about the channel does not get outdated at the NodeB. It is shown that
the maximum prediction horizon at 120 km/h corresponds to a covered distance of one half of
the wavelength. Using Slepian sequences for channel estimation as well as prediction in timedomain is similar to the reduced rank Wiener interpolation and prediction respectively with uniform Doppler spectrum assumption, since the subspaces spanned by both estimators are exactly
the same. The performance of the presented methods for time-varying channel estimation has been
evaluated with the SCME channel model. The 2.5 MHz and 20 MHz downlink transmission bandwidths within the LTE have been investigated. It has been shown that the LMMSE CIR estimator
with the Slepian sequence in general has the MSE performance, since it is closest to the Wiener
interpolator. Hence the LMMSE CIR estimator is a good compromise between complexity and
performance.
Future work
This thesis investigated channel estimators for the UMTS LTE downlink structure with transmission bandwidths 2.5 MHz and 20 MHz. For future research with other transmission bandwidths the
implemented Matlab simulator can easily be recongured to evaluate the estimation performance.
We have not considered coding in this thesis, hence this is left as future work. The different lowcomplexity estimators developed in this project can be used with the applied coding in order to
nd an appropriate estimator. Moreover the evaluation of the estimators shows that if the speed of
the UE is below the speed at which the Slepian sequences are adapted, the performance decreases
with respect to the Wiener interpolator. Hence possible future work is also to estimate the speed of
the UE in order to improve the channel estimation as well as prediction. This work can be based
on [27] which already presents an approach for this task.
72
Bibliography
[1] 3GPP. 3GPP TR 25.913 V7.3.0 Requirements for Evolved UTRA (E-UTRA) and Evolved
UTRAN (E-UTRAN) . Technical report, 3GPP, September 2003.
[2] 3GPP. 3GPP TR 25.996 V6.1.0 Spatial Channel Model for Multiple Input Multiple Output
(MIMO) simulations (Release 6). Technical report, 3GPP, September 2003.
[3] 3GPP. 3GPP TR 25.943 V6.0.0 Deployment Aspects (Release 6). Technical report, 3GPP,
December 2004.
[4] 3GPP. 3GPP TR 25.814 V7.1.0 Physical Layer Aspects for Evolved Universal Terrestrial
Radio Access (UTRA) (Release 7). Technical report, 3GPP, September 2006.
[5] 3GPP. 3GPP TS 36.211 V0.3.1 Physical Channels and Modulation (Release 8). Technical
report, 3GPP, February 2007.
[6] Andrea Ancora, Calogero Bona, and Dirk T M Slock. Down-sampled Impulse Response
Least-Squares Channel Estimation for LTE OFDMA. In ICASSP 2007, 32nd IEEE International Conference on Acoustics, Speech, and Signal Processing, Honolulu, USA, Apr 2007.
[7] P. Dent, G.E. Bottomley, and T. Croft. Jakes fading model revisited. In Electronics Letters,
volume 29, pages 11621163, 24 June 1993.
[8] O. Edfors, M. Sandell, J. van de B., B. Wilson, and P. Brjesson. OFDM Channel Estimation
by Singular Value Decomposition. In IEEE Vehicular Technology Conference, Atlanta, USA,
pp. 923-927,April 28-May 1., 1996.
[9] B.H. Fleury. An Uncertainty Relation for WSS Processes and Its Application to WSSUS
Systems. IEEE Transactions on Communications, 44(12):16321634, Dec. 1996.
[10] G. H. Golub and C. F. Van Loan. Matrix Computations. Baltimore, MD, USA, second
edition, 1989.
[11] S Haykin. Adaptive Filter Theory. Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 1995.
[12] Harri Holma and Antti Toskala. WCDMA for UMTS. John Wiley & Sons, Inc., New York,
NY, USA, 2002.
BIBLIOGRAPHY
[13] Stefan Kaiser. Multi-Carrier CDMA Mobile Radio Systems Analysis and Optimization of
Detection, Decoding, and Channel Estimation. PhD thesis, University of Kaiserslautern,
Germany, January 1998. VDI Verlag Dsseldorf, Series 10, No. 531, ISBN 3-18-353110-0.
[14] Steven M. Kay. Fundamentals of Statistical Signal Processing: Estimation Theory. PrenticeHall, Inc., Upper Saddle River, NJ, USA, 1993.
[15] Louis Litwin and Michael Pugel. The principles of OFDM, January 2001.
[16] Andrew M. Odlyzko. Internet trafc growth: Sources and implications, 2003.
[17] Alan V. Oppenheim and Ronald W. Schafer. Digital Signal Processing. PrenticeHall, 1989.
[18] M.K. Ozdemir, H. Arslan, and E. Arvas. Adaptive Low-Rank MIMO-OFDM Channel Estimation. In Wireless and Microwave Technology, 2005. WAMICON 2005. The 2005 IEEE
Annual Conference, page 5pp., 2005.
[19] S. Parkvall, E. Englund, M. Lundevall, and J. Torsner. Evolving 3G Mobile Systems: Broadband and Broadcast Services in WCDMA. IEEE Communications Magazine, 44(2):3036,
Feb. 2006.
[20] A. Pokhariyal, G. Monghal, K. I. Pedersen, P. E. Mogensen, I. Z. Kovacs, C. Rosa, and T. E.
Kolding. Frequency Domain Packet Scheduling Under Fractional Load for the UTRAN LTE
Downlink. In Vehicular Technology Conference, 2007. VTC2007-Spring. IEEE 65th, pages
699703, April 2007.
[21] Jari Salo, Giovanni Del Galdo, Jussi Salmi, Pekka Kysti, Marko Milojevic, Daniela Laselva,
and Christian Schneider. MATLAB implementation of the 3GPP Spatial Channel Model
(3GPP TR 25.996). On-line, January 2005. http://www.tkk./Units/Radio/scm/.
[22] D. Slepian. Prolate spheroidal wave functions, Fourier analysis, and uncertainty. V - The
discrete case. AT&T Technical Journal, 57:13711430, June 1978.
[23] Antti Toskala, Harri Holma, Kari Pajukoski, and Esa Tiirola. Utran Long Term Evolution
in 3GPP. In IEEE 17th International Symposium on Personal, Indoor and Mobile Radio
Communications, pages 15, Sept. 2006.
[24] J.-J. van de Beek, O. Edfors, M. Sandell, S.K. Wilson, and P.O. Borjesson. On Channel
Estimation in OFDM Systems. In IEEE 45th Vehicular Technology Conference, volume 2,
pages 815819vol.2, 25-28 July 1995.
[25] T. Zemen, H. Hofstetter, and G. Steinbock. Successive Slepian Subspace Projection in Time
and Frequency for Time-Variant Channel Estimation, 2005.
[26] T. Zemen and C.F. Mecklenbrauker. Time-Variant Channel Estimation Using Discrete Prolate Spheroidal Sequences. IEEE Transactions on Signal Processing,, 53(9):35973607,
Sept. 2005.
74
BIBLIOGRAPHY
[27] T. Zemen and Fleury B. H. Mecklenbrauker, Kaltenberger F. Minimum-Energy BandLimited Predictor with Dynamic Subspace Selection for Time-Variant Flat-Fading Channels.
IEEE Transactions on Signal Processing, 2007. Accepted (2006-12-06), to be published.
[28] Thomas Zemen. OFDM Multi-User Communication Over Time-Variant Channels. PhD
thesis, 2004. ISBN 3-902477-04-0.
75
Multiplexing systems
In high data rate, wideband communication through multipath fading channel, the received signals
are affected by inter-symbol interference (ISI) due to the fact that the time spread of the channel is
relatively large compared to the symbol duration. Reducing ISI is an important issue for enhancing
the quality of wideband communications.
Normally ISI is removed by channel equalization, which is implemented either in the time or
frequency domain, with symbol-by-symbol or sequence estimation algorithms.
If the symbol rate is reduced, while the information rate is remained, the impact of ISI becomes
negligible and equalization becomes unnecessary. One method is to increase the constellation size
in an Q-ary pulse modulation scheme. However, there is a limit on how large Q can be before
modulation and demodulation becomes too complex.
Multicarrier modulation offers another approach. A serial high rate data stream is distributed into
many parallel low rate data streams, transmitted through many orthogonal narrowband subcarriers,
thus the symbol duration in each channel is large enough that the ISI can be negligible.
In this Appendix, the principle of multicarrier modulation is reviewed, followed by a review of
OFDM, which is based on [15].
Figure A.1: Basic structure of multicarrier modulation system. PS is the pulse shaping filter at the
transmitter side, while MF is the matched filter at receiver side.
fs
NIFFT
(A.1)
A.3 Orthogonality
An OFDM system is shown in Figure A.2. The Fast Fourier Transform (FFT) algorithms are used
to generate orthogonal subcarriers due to their computational efciency [15]. The IFFT operation
results in the composite OFDM symbol of duration Ts . The idea behind using IFFT to modulate
78
A.3. ORTHOGONALITY
Figure A.2: Basic structure of an OFDM system. PS is the pulse shaping filter at the transmitter side
while MF is the matched filter at the receiver side.
(A.2)
where sn (t) is the nth modulated subcarrier, An (t) and n (t) are the modulated amplitude and
phase of the nth subcarrier respectively, f is the frequency spacing between subcarriers, i.e. subcarriers are a group of harmonic sinusoids. The composite OFDM signal is then the summation of
all NIFFT subcarriers [15].
S(t) =
NIFFT
1
2
NIFFT
1
2
s(t) =
N
n= IFFT
2
(A.3)
N
n= IFFT
2
where N is an even number of subcarriers and S(t) is a baseband signal that is subsequently
upconverted to the desired carrier frequency before transmission. Now sampling S(t) over one
OFDM symbol period Ts , with a sampling frequency of fs =
NIFFT
1
2
S(kTC ) =
1
TC ,
An ej(2f knTC +)
results in [15]
(A.4)
N
n= IFFT
2
where Ts = N TC and the amplitude and phase of each subcarrier are constant in the duration of
S(kTC ) =
NIFFT
1
2
j N2 f knTs
An ej e
IFFT
(A.5)
N
n= IFFT
2
79
S(kTC ) =
j N2 kn
An ej e
(A.6)
IFFT
N
n= IFFT
2
1
Ts
orthogonal. Orthogonality is clear from examining the spectrum of the OFDM signal in Figure
A.3. The IFFT block samples the spectrum at the peak of each sinc function to map An and n to
subcarriers. This results in no interference between subcarriers. The overlapping spectra is not of
importance because the peak of each sinc functions coincides with zero crossings of all other sinc
functions. In the time domain, orthogonality means that all subcarriers have an integer multiple of
periods in each symbol period Ts as shown in Figure A.4.
1.2
subcarrier 1
1
0.8
0.6
0.4
0.2
0
f
0.2
0.4
10
Figure A.3: The spectrum of an OFDM signal consists of overlapping sinc functions. Each sinc function is
sampled at its peak, which coincides with zero crossings of all other functions [15].
Because of the nature of a mobile environment, multiple delays of the transmitted signal reach a
mobile receiver. Delay spread results in a frequency selective fading channel response. Hence,
some signal frequencies are enhanced while others suffer from deep fades. This affects both the
amplitude and phase modulation of the carrier and consequently results in information loss. In
order to combat frequency selective fading, OFDM uses a signal bandwidth that is much smaller
than the coherence bandwidth associated with the estimated delay spread of the channel. This
80
1
0.8
0.6
0.4
0.2
0
t
0.2
0.4
0.6
0.8
1
50
100
150
200
250
300
Figure A.4: Orthogonality in the time domain means that all subcarriers in an OFDM signal have an
integer number of periods during the period of one OFDM symbol [15].
way, each subcarrier experiences fading and only some bits are lost if some of the subcarriers
experience deep fades [15].
Because an OFDM signal is composed of multiple subcarriers that are independently modulated,
the amplitude of the signals envelope can reach high peaks if all subcarriers add coherently. When
compared to the envelopes average amplitude level, such peaks result in a high Peak-to-Average
Power Ratio (PAPR) [15]. An OFDM signal composed of NIFFT subcarriers can have a PAPR
equal to 10 log N [15].
81
1.5
0.5
0.5
1
cyclic prefix
1.5
TC
TCP 5
10
15
Ts 20
25
30
Figure A.5: A cyclic prefix CP is added before each OFDM symbol [15]. This achieves continuous transmission of all subcarriers required for transmitter/receiver synchronization and makes it possible to use
simple frequency domain equalization.
A.6.2
An OFDM receivers ability to separate the overlapping subcarriers of an OFDM signal relies
on orthogonality of these subcarriers. In an OFDM system, frequency and time synchronization
between the transmitter and the receiver are crucial to maintain orthogonality of the subcarriers.
Loss of orthogonality results in inter carrier interference (ICI) and ISI [15]. The effect of imperfect
frequency synchronization between the transmitter and receiver in an OFDM system can be shown
by examining (A.5) once again [15],
NIFFT
1
2
S(kTC ) =
j N2 kn
An ej e
IFFT
(A.7)
N
n= IFFT
2
As mentioned in Section A.3, S(kTC ) is the baseband discrete-time OFDM signal. The upconverted transmitted signal, SUC (t), is then the analog version of S(kTC ). The analog version is a
result of a convolution between pulse shaping lter and the discrete signal, which is also depicted
on Figure A.2.
S(t) is upconverted to the desired carrier frequency fc [15],
SUC (t) = S(t)ej2fc t
(A.8)
Now, at the receiver side, SUC (t) should be downconverted to baseband by multiplying the signal with the term ej2fc t to obtain S(t) once again. However, if the receivers local oscillator
82
(A.9)
where foff is the receivers frequency error due to a frequency shift. The term ej2foff t accounts
for ICI occurring due to the receivers frequency error. In other words, subcarriers no longer have
integer number of periods during an OFDM symbol period. Now, assuming a perfect matched
ltering and sampling frequency at the receiver, S(t) results in a discrete time signal, S(kTC ), in
the form [15]
Soffset (kTC ) = S(kTC )ej2foff kTC =
NIFFT
1
2
j N2 kn
IFFT
(A.10)
N
n= IFFT
2
The receivers FFT block then performs digital demodulation of Soffset (kTC ) to restore the transmitted constellation points. In effect, this is equivalent to multiplying S(kTC ) with the term
j N2 kn
IFFT
. Therefore, when the distorted signal Soffset (kTC ) is applied to the FFT block, the re-
Sc (k) =
(A.11)
N
n= IFFT
2
where Sc (k) are the constellation points obtained at the output of the FFT block.
Hence, the effect of imperfect frequency synchronization is a phase rotation of constellation points
by an angle error = 2foff kTC . It can be seen that err increases with both the frequency error and
time.
Loss of orthogonality can also result from loss of timing synchronization. At the receiver side, the
boundaries of OFDM symbols have to be located in order to apply the FFT window at the right
time.
Figure A.6: When there is a time shift, ISI occurs and hence received constellation points are distorted.
[m] = TCP
1
F H d[m].
NIFFT NIFFT
(A.12)
The discrete time has the rate Ts and after parallel to serial conversion of [m], the signal is
transmitted over a multipath channel.
The NIFFT -point DFT-matrix FNIFFT CNIFFT NIFFT with elements
j2il
i, l {0, 1, . . . NIFFT 1} .
(A.13)
"
0NIFFT LCP
ICP
INIFFT
RP NIFFT .
(A.14)
Matrix TCP copies the last LCP chips of each OFDM symbol to the front.
If [n] denotes the chip sequence with rate 1/Tc to be transmitted, then
[m] =
[mP ]
..
.
[mP + P 1]
CP
(A.15)
h =
h [0]
..
.
h [L 1]
CL
(A.16)
L
X
k=0
h[k][n k].
(A.17)
The received signal r is also affected by white Gaussian noise z[n] (N )(0, z2 ),
r[n] = x[n] + z[n].
(A.18)
The elements x[n] and z[n] has the following structure for each OFDM symbol,
x[m] =
84
x[mP ]
..
.
x[mP + P 1]
CP
(A.19)
and
z[m] =
z[mP ]
..
.
z[mP + P 1]
CP .
(A.20)
The inuence of the channel is included by taking the lower and upper triangular Toeplitz channel
matrix into account.
H (0)
and
h [0]
..
.
0
..
.
0
..
.
h [L 1]
=
..
.
0
.
.
..
..
0
0
. . . 0 h [L 1] . . . h [0]
H (1)
h [1]
0 . . . 0 h [L 1] . . .
..
..
..
..
.
.
.
h [L 1]
=
0
0
.
..
..
.
0 ...
CP P
(A.21)
CP P
(A.22)
are the upper and lower triangular Toeplitz channel matrix respectively.
Now the intersymbol interference is modeled as,
x[m] = H (0) [m] + H (1) [m 1],
(A.23)
where the second term represents the inter-symbol interference between two consecutive OFDM
symbols, which affect the rst L samples of x[m].
At the receiver the cyclic prex of length LCP is removed by forming the product
RCP H (1) = 0NIFFT P ,
where
(A.24)
(A.25)
with 0NIFFT P and 0NIFFT LCP as matrices containing zeros with the dimensions NIFFT P and
NIFFT LCP respectively. Equation (A.24) only holds if the maximum channel delay L 1 is
less than or equal to the cyclic prex LCP . The received signal is now written as,
(A.26)
85
(A.27)
(A.28)
with FL as the truncated version of F with the rst L columns. H (0) has a Toeplitz structure and
by inserting the cyclic prex it is circularized. Hence the DFT-matrix are the eigenvectors, while
the channel matrix diag(g) is diagonal. Using Equation (A.26) and (A.27), the received signal can
be expressed as,
y[m] = diag(g)d[m] + z[m].
(A.29)
The elements of the noise vector, z[m] are white Gaussian with variance z2 and the covariance
matrix of z[m] is diagonal with identical values
H
Rz[m] = E F RCP z[m]z[m]H RCP
FH
=
H
z2 F RCP RCP
FH
= z2 INIFFT .
(A.30)
(A.31)
(A.32)
From Equation (A.29), each symbol vector d[m] is transmitted over an individual frequency-at
subcarrier.
A.8 Summary
Advances in DSPs have generated a great interest in OFDM as a spectrally efcient MCM scheme,
that is capable of delivering high data rates. The FFT is the basic algorithm in any OFDM system, because it allows for a computationally efcient implementation. OFDM has high immunity
against frequency selective fading and ISI which makes it suitable for mobile communication.
OFDM transmission is also very sensitive to frequency and timing errors between the transmitter
and the receiver, which result in loss of orthogonality and therefore, ICI and ISI.
86
As described in Section 2.4.2 the reference signal sequence Rref is generated as a symbol-bysymbol product of an orthogonal sequence ROS C3402 and a pseudo-random sequence RPRS
R3402 . If rm,n is the entry at mth row and nth column of Rref , where m = 0..339 and n = 0..1,
then
OS
PRS
rm,n = rm,n
rm,n
,
(B.1)
There are NOS = 3 different orthognal sequences and NPRS = 170 different psudo-random sequences. The cells are divided into three identity groups at the physical layer level. There is a
one-to-one mapping between the three identities and the three two-dimensional orthogonal sequences.
(B.2)
where sm,n is the entry at the mth row and the nth column of the matrix Si , dened as
(B.3)
114 repetitions
where
1 1
ej4/3
ej2/3
j2/3
j4/3
S1T =
1
1
1 1 S1 = e
S1 = e
j4/3
j2/3
j2/3
j4/3
1 1
e
e
e
e
(B.4)
symbols (ak,l (i)), where k is the OFDM symbol position in one slot and l is the position of the
subcarrier, while p and i are the antenna port and slot number respectively:
(p)
al,k (i) = rm ,n ,
(B.5)
m = m + 170
NRB
m = 0, 1, . . . ,
n=
0 and 1 if p = 0 or p = 1
0
if p = 2 or p = 3
(B.6)
(B.7)
(B.8)
The entry (l, k) of the reference symbol sequence is then calculated as following,
k = 6m + (v + fhop (i/2)) mod 6 where
l=
B.4.1
(B.9)
if n = 0 and p = 0 or p = 1 (1)
1
if n = 0 and p = 2 or p = 3 (2) .
N
sym 3 if n = 1 (3)
(B.10)
In this example the number of transmitter antennas is 1 (p = 0), in order to make it as simple
as possible. Using transmission bandwidth of 2.5 MHz, NBW = 150 according to Table 2.1 and
NRB = 12 as explained in section . The number of OFDM symbols in one slot Nsym = 6 and it is
chosen not to use frequency hopping.
The reference symbol to be transmitted is found in the following procedure:
- The index (m , n) is calculated.
Since p = 1, n = 0 and n = 1 (Equation (B.8)).
m = m = 0, 1, . . . , 150
6/2 1 = 0, 1, . . . , 24 (Equation (B.6)).
BW
m = m + 170 N
NRB = 158, 159, . . . , 182 (Equation (B.7)).
88
Figure B.1: Example of the reference symbol structure for one slot with 6 OFDM symbols using one
antenna. Note that only the used subcarriers are depicted.
89